Telmi · Accounting intelligence
Tell me — before I ask
Know what’s happening.
WhatsApp par apna Tally. Ask anything in Hindi, Hinglish or English — better, Telmi tells you first: low cash, overdue parties, duplicate payments, 34 automations watching your books. And it physically cannot touch them.
A Retrogrades product · Multi-tenant SaaS · Free starter tier · Business ₹1,499/mo
The world before
One number. One desk-bound PC. One phone call to the accountant.
To learn what a party owes, today’s collections, or cash in hand, the owner must be at the accounting PC — or phone the accountant and wait. And the problems that hurt most — a cash crunch, an overdue receivable, a duplicate payment — surface after they’ve already cost money.
The insight
Don’t teach the owner new software. Meet him where he already is.
No new app. No dashboard. No change to the accounting licence. A small agent reads Tally read-only and syncs to the cloud; the owner WhatsApps “Sharma Traders ko kitna payment hua is mahine?” — and gets the answer, a tapped menu report, or a rendered PDF statement back in the chat. In the language he actually speaks.
The product
A real accounting engine on the other end of the chat.
Sixty tenant-scoped computations
Balance sheet, P&L, trial balance, GST position, bill-wise outstanding with ageing, cash runway, project profitability — computed from Tally’s own rolled-up balances so figures match it to the rupee.
Hinglish, understood deterministically
A deterministic Hindi/Hinglish parser handles ~80% of queries with no AI at all — 14 intents, Devanagari keywords, a WhatsApp-typo normaliser and Indian financial-year dates. “Kitna lena hai” means receivables. It knows.
PDFs, delivered into the chat
Ledger and party statements, business snapshots, balance sheets, P&L and tax invoices — rendered with Devanagari support and dropped straight into WhatsApp.
34 automations that watch the books
Low cash, overdue invoices, duplicate payments detected by clustering with confidence scores, sales drops, silent customers, expense spikes, month-end P&L, tax reminders — pushed before problems hurt, not after.
The difference
“Can this mess up my books?”
No. Architecturally, no.
The connector cannot write. The agent on your PC is read-only with a hard write-guard: any instruction other than an export request raises an error before it leaves the machine. Your machine opens no inbound port.
The AI cannot improvise. When the deterministic parser hands over to a model, that model can only choose from twelve whitelisted tools. It never writes SQL, never sees your tenant identifier, and receives only the result of a query to phrase — with every number formatted server-side before the model sees it, and recomputation forbidden.
The product checks itself. Every proactive alert’s numbers are re-derived through a separate code path before sending. An autonomous QA loop re-computes production answers against ground truth every hour, fixes root causes, and mandates a regression test for each one.
Strangers get nothing. Only verified numbers bound to your tenant can ask. Unknown numbers get a registration pointer — never data.
Telmi reads your books; it never writes them. It surfaces GST positions and tax reminders as estimates for awareness — it is not filing or compliance software.
The product, for real
The console behind the chat.
Live screens from a real tenant — figures softened. The WhatsApp thread is the front door; this is the intelligence behind it.
The business snapshot — cash position, runway and who to call today, in plain language
Receivables & payables — kaun paisa dega, kis ka dena hai
Real questions, answered — Hinglish queries with routing and accuracy tracking
Your books, in your pocket, in your language.
Pair the agent with a 6-digit code, and ask your first question five minutes later.